ANALYSIS The enterprise AI conversation has pivoted. The question keeping CIOs up at night is no longer which foundation model to deploy but how to govern, meter, and control the multi-agent systems already running in production — and the data suggests most organizations are not yet equipped to do so.
Why it matters
Agentic AI is moving from experimentation into production, and enterprises are discovering that the hard problems are infrastructure problems1. Across 107 enterprises surveyed in a July 2026 VentureBeat Pulse Research wave, 85% run two or more orchestration platforms and 64% run three or more, with a mean of 3.1 platforms per organization2. That fragmentation is not a bug — it reflects a deliberate strategy to maintain flexibility across models — but it is creating a governance and cost-control vacuum that few have filled. One in five enterprises still has no real-time, programmatic way to stop a runaway agent before a budget-breaking bill arrives.
The big picture
The survey data paints a market where platform incumbency and forward momentum are splitting apart. Microsoft AI Foundry / Copilot Studio appears in 70% of enterprise stacks and OpenAI's Agents SDK in 68%, with Anthropic's Claude Platform in 47%. But when the 72 enterprises actively evaluating new or replacement orchestration platforms were asked about forward consideration, Anthropic led at 43%, followed by Google and custom in-house builds each at 31%, OpenAI at 25%, and Microsoft and LangChain / LangGraph each at 17%. Sixty-seven percent of enterprises intend to adopt a new, additional, or replacement orchestration platform within the year.
ANALYSIS The gap between current deployment share and forward consideration suggests that enterprises are not satisfied with the control planes their incumbent providers offer — and are shopping for alternatives that prioritize governance and cost transparency over ecosystem convenience.
The risk enterprises fear most from provider-resident control is not lock-in but the provider's own security and permissioning limits. That finding aligns with a separate governance framework from Liferay, which argues that organizations need six elements in place before deploying AI agents: access control, a single source of truth, audit trails, human review checkpoints, a clear escalation path, and model-agnostic architecture3. The framework's emphasis on model-agnostic architecture — building controls at the platform layer, independent of any one model — echoes the broader industry shift from model selection to platform control.
Between the lines
The most revealing data point in the VentureBeat survey may be how little of the deployed agent estate is genuinely orchestrated. Only 16% of enterprises say more than half of their agents are genuinely orchestrated. Thirty-seven percent say a quarter or fewer are. The plurality — 47% — report that between 26% and 50% of their agents meet that bar.
ANALYSIS In other words, the majority of what enterprises call "agents" today are operating with limited multi-step coordination, which helps explain why task completion reliability was the top orchestration success metric at 30%, followed by multi-step workflow management at 27%. Enterprises are measuring what they lack.
Cost governance remains primitive. Only 30% of enterprises rely entirely on native caps and throttles built into their primary platform for fiscal control. Twenty-five percent use custom gateways, and 24% use cross-model routing to arbitrage cost. The remaining 21% with no real-time cost controls represent a material operational risk as agent volumes scale.
Singh Sanatya, an enterprise technology leader and IEEE author, frames the challenge in architectural terms: successful transformation "isn't about implementing technology — it's about creating sustainable organizational capability"4. That language mirrors what the Liferay framework makes concrete: an AI agent "does not reconcile conflicting information unless it has been trained to do so" and "delivers whatever it can access" with full confidence, even when underlying data is wrong. Without governed data pipelines and escalation paths, scaling agents scales errors.
Security and permissions enforcement accounts for 30% of planned orchestration-related spend growth, and combined with monitoring and debugging, the two categories account for 61% of planned growth. ANALYSIS Enterprises are signaling that the next dollar of agentic AI investment goes not to model access but to the control plane around it.
What's next
With 67% of enterprises planning to adopt a new orchestration platform within the year, the competitive landscape is fluid. Anthropic's lead in forward consideration at 43% — nearly double OpenAI's 25% — sets up a contest for the emerging platform layer that will play out through procurement cycles already underway. The 46 of 107 respondents who could not name a single primary platform underscore how unsettled the market remains. ◆ The enterprises that solve metering and governance first will not just control costs — they will define the architectural standard for production agentic AI.